Domain 8 of 10

Neural Potentials, from Your Own Data

Neural network potentials trained only on engine data you computed, with ensembles, active learning and sealed models.

2 Partial2 UnverifiedEngine 1.10.12 live

A neural network feeding a potential energy surface

What You Can Compute

Your Data Only

Sampling runs your chosen quantum method, or you import your own extended-XYZ or JSON data (since 1.9.3). Other groups' data and models are never read.

Training

Five architectures, energy and force training, committee ensembles and active learning.

Sealed Models

Data sets and models are sealed with SHA-256, so every prediction names its model.

Every Method, with Its Level of Support

Generated from the engine's own capability report (engine 1.10.12), so this list never runs ahead of the code.

Checked reproduces printed values Checked (exact) no printed value exists; checked against exact mathematics or brute force Partial code exists, unchecked parts badged in the app Unverified built, no printed check yet

Method	Support	Since	Primary Sources
Neural network potentialsBehler-Parrinello potentials with ANI-type descriptors and analytic forces. Since 1.9.3 you can import your own extended-XYZ or JSON reference data; units, finite values, distances, forces, duplicates and energy outliers are checked and the data set is sealed. Trained only on your own data (engine-generated or imported by you); models stay in your organisation and are labelled trained on user-provided data.	Unverified	0.9.0	Behler, Parrinello 2007; Smith, Isayev, Roitberg 2017
Force training, ensembles, active learningEnergy and force training, committee ensembles and uncertainty-driven sampling.	Partial	1.1.0	Smith et al. 2018
MLP dynamics and optimisationUse a trained model in MD and geometry optimisation.	Partial	0.9.0
SchNet, NequIP-style, Deep Potential and GAPMessage-passing, equivariant, local-frame and SOAP kernel potentials trained on your own engine data. Not periodic, no GPU training.	Unverified	1.5.0	Bartok, Csanyi 2015; Zhang et al. 2018

Printed Value, Our Value

No printed-value rows are listed for this domain yet. Self-tests and cross-checks are described on the Validation page.

All validation rows

Not Offered Today

Pretrained foundation models
Data from other groups, public data sets or pretrained weights (your own imported data is accepted)
Periodic MLP dynamics

In the Chemistry Lab

Start, then Large systems and dynamics, then Train a potential. The flow samples, trains, checks and then lets you use the model in dynamics or optimisation.

Machine-Learned Potentials Questions

Can I train on my own reference data?

Yes. The import tool reads extended-XYZ or JSON written by you, with a declared level of theory and units. It checks units, finite values, distances, force completeness and size, net force, duplicates and energy outliers, refuses mixed levels of theory unless you allow them, and seals the data set. Every model trained on it is labelled trained on user-provided data and stays in your organisation.

Why only our own data?

So every model can be traced to computations we can check. A model with unknown training data cannot be validated.

How do I know a model is good enough?

The check tool reports errors on held-out data and the ensemble spread; active learning adds points where the spread is large.

See It on a Molecule of Your Own

Pair your own server once, then let your team queue computations from the browser.

Get Started Request a Demo